Gemini 3.7 Flash deprecated: legacy API routes to 3.8 Flash

Checked October 11, 2026: Google deprecated Gemini 3.7 Flash on October 8, 2026. Calls to gemini-3.7-flash now automatically run gemini-3.8-flash. No final shutdown date has been announced, but the old ID no longer serves an independent 3.7 model. The 3.7 benchmarks below are historical, not new API measurements. Google release notes · model page. ¹²

Gemini 3.7 Flash specifications

Item	Gemini 3.7 Flash
Model ID	gemini-3.7-flash (legacy) → gemini-3.8-flash
Status	deprecated, Oct 8, 2026; routes to 3.8
Input limit	1,048,576 tokens
Output limit	65,536 tokens
Inputs	text, image, video, audio, PDF
Output	text
Thinking	low, medium, high
Computer use	Preview
Function calling	supported
Structured output	supported
Google Search / Maps grounding	supported
Live API	not supported

For the historical 3.7 specification: minimal thinking is not supported and returns an error according to Google. ¹

Deprecation: 3.7 vs 3.8 is no longer a valid live A/B test

Google distinguishes deprecation from permanent shutdown. However, the model page confirms the 3.7 alias already routes to 3.8 Flash. An A/B test between these two IDs therefore calls the same current successor twice. To compare distinct active models, use 3.6 versus 3.8 Flash where available. The original 3.7 benchmarks remain useful only as historical vendor data. ¹²

API pricing through 2026 and from 2027

Google lists the following paid Standard rates for 3.8 Flash, the model now serving the old alias:

Period	Input per 1M tokens	Output per 1M tokens
through Dec 31, 2026	$0.75	$3.75
from Jan 1, 2027	$1.50	$7.50

These are 3.8 prices, not an active standalone 3.7 tariff. Output pricing includes thinking tokens. Context caching is $0.075 per million cached input tokens through 2026 and $0.15 from January 2027. Cache storage is billed separately per token and hour. ³

What 3.7 improved over 3.6

Google’s launch-era model-card evaluations showed substantial gains over 3.6 on several coding and agent tests, including DeepSWE, Terminal-Bench 2.1, Automation and MRCR 128K. Those results explain why 3.7 was a meaningful August upgrade, but they are vendor evaluations rather than independent Mac benchmarks.

Historical model evidence: The 3.7 scores above are vendor data from launch, not a current A/B test against 3.8. The old 3.7 ID now routes to 3.8 and cannot serve as an independent control. Evaluate agents by task quality, tool errors, latency and cost per completed task. ¹²

Calling active Gemini 3.8 Flash from a Mac

Gemini 3.8 Flash is not a local Apple Silicon model. Calls to the old 3.7 alias also run as 3.8 on Google’s infrastructure. The Mac runs the client, SDK, editor, terminal and any local preprocessing. ¹

Before running Python, install the Google GenAI SDK and set your API key in the GEMINI_API_KEY environment variable. The SDK reads it automatically. Do not put credentials in a repository or client-side code; see Google’s API-key guide.

python3 -m pip install -U google-genai

This example uses the supported Generate Content interface. Google’s newer Interactions API uses different request and response fields, so snippets from the two APIs are not interchangeable.

A minimal Python call:

from google import genai

client = genai.Client()

response = client.models.generate_content(
    model="gemini-3.8-flash",
    contents="Review this function for race conditions."
)

print(response.text)

The 3.8 ID selects the active model. The old 3.7 ID is not a rollback because it also routes to 3.8. ¹⁴

Comparing active 3.6 and 3.8 Flash APIs

Avoid a false benchmark: Since gemini-3.7-flash routes to 3.8, this example explicitly compares 3.6 with 3.8 rather than calling 3.8 twice. Both are cloud models. This does not recreate a historical 3.7 benchmark. Validate model-specific parameters. Google migration guide.

Prerequisites: Python 3, google-genai, and a GEMINI_API_KEY created in Google AI Studio. These are billable cloud requests where paid access is used; use a deliberate spending cap. Do not send passwords, client data or private repository contents as examples.

python3 -m pip install -U google-genai
read -rs GEMINI_API_KEY
export GEMINI_API_KEY

Save the following as compare_gemini_flash.py. It submits the same synthetic task to each model and prints response timing and Google’s token-usage fields without printing your API key or the generated content.

import os
import time
from google import genai

if not os.environ.get('GEMINI_API_KEY'):
    raise SystemExit('Set GEMINI_API_KEY before the comparison')

client = genai.Client()
prompt = 'List three specific criteria for reviewing a Python pull request.'

for model in ('gemini-3.6-flash', 'gemini-3.8-flash'):
    started = time.perf_counter()
    try:
        response = client.models.generate_content(model=model, contents=prompt)
    except Exception as error:
        print(f'{model}: request failed ({type(error).__name__})')
        continue
    elapsed = time.perf_counter() - started
    usage = response.usage_metadata
    print({
        'model': model,
        'elapsed_seconds': round(elapsed, 2),
        'prompt_tokens': getattr(usage, 'prompt_token_count', None),
        'answer_tokens': getattr(usage, 'candidates_token_count', None),
        'thinking_tokens': getattr(usage, 'thoughts_token_count', None),
    })

Read the output correctly: elapsed_seconds measures one cloud request’s wall-clock latency, not Apple Silicon inference speed. Google’s token accounting guide says prompt_token_count includes cached input tokens; thoughts_token_count represents thinking tokens, separate from visible answer tokens. Total API cost can also depend on grounding, caching, additional tool calls and platform rates. For a useful comparison, repeat at least three times across five appropriately sanitized real-world tasks, log the chosen thinking level and report aggregate accuracy, time and cost per accepted result.

Production migration checklist

• Model identifier: Switch gemini-3.7-flash to gemini-3.8-flash only in a separately tested candidate deployment.
• Sampling: Per Google’s current checklist, remove legacy temperature, top_p, top_k, candidate_count and prefilled model turns rather than forwarding them unexamined.
• Thinking: Migrate thinking_budget to thinking_level: low, medium or high. minimal is unsupported in 3.8 and causes an error. Keep the chosen level fixed during comparisons.
• Multi-turn agents: With the Interactions API, use previous_interaction_id where required. Its message and tool semantics are different from the generate_content example above; validate tool responses and IDs separately.
• Rollback: Test correctness, cost and tool failures. The old 3.7 ID is not a real rollback because it already routes to 3.8; pre-validate another active fallback.

Primary references: Gemini 3.8 migration guide, token usage fields, deprecation schedule. Our documentation presents verified API schemas without asserting live execution figures.

Migrating existing legacy integrations

An old 3.7 ID is not a version pin. Existing integrations already receive 3.8 responses even if their configuration still says 3.7. Set gemini-3.8-flash explicitly and retest structured output, tools and thinking. Historical 3.7 comparisons require evidence dated before October 8, 2026.

Current role of Gemini 3.7 Flash

Gemini 3.7 Flash is deprecated and routes to Gemini 3.8 Flash. Current pricing and newly generated API results refer to 3.8, not an independently callable 3.7. Compare active 3.6 and 3.8 for live testing. No independent API or Mac benchmarks were run.

Sources and references

Model limits, pricing and retirement status come from Google’s own documentation. Model benchmark claims are vendor-reported, not our own Mac measurements.

Footnotes

1. Google – Gemini 3.7 Flash model ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

2. Google – Gemini API changelog ↩ ↩² ↩³

3. Google – Gemini API pricing ↩

4. Google – Gemini 3.8 Flash model ↩